alokkrsharma wrote:

> There are unanswered questions in 
> https://discourse.llvm.org/t/rfc-pass-instrumentation-based-verification-for-pgo-use-profile-consistency/90796?u=boomanaiden154-1
>  and it is still unclear how this integrates with profcheck.

Thanks, I agree there are still open questions on the RFC thread, especially 
around scope and the relationship with ProfCheck.

Just to clarify, I'm not expecting this PR to land before those discussions are 
resolved. The current branch is primarily a PoC to validate the RFC ideas on 
real optimization pipelines and to measure the signal/noise characteristics of 
the approach.

Since the RFC was posted, I've made a number of changes aimed at reducing false 
positives (e.g. the new approxprofile attribute, opt-in aggressive checks, 
indirect-call handling, diagnostic deduplication, and a standalone verifier 
mode in addition to the instrumentation-based mode).

Regarding ProfCheck, I currently see them as complementary. One goal of the 
instrumentation-based mode is to help identify which pass first introduces a 
profile-flow inconsistency. The standalone mode is more about validating 
whether flow-conservation properties still hold after a sequence of 
transformations.

Once the RFC direction is clearer, I'm happy to split this work into smaller 
pieces and revisit what should remain separate versus what could potentially 
build on or integrate with existing ProfCheck infrastructure.

https://github.com/llvm/llvm-project/pull/196295
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to